Event date · · Z.ai

GLM-5.2 Open Source: Zhipu Uses 1M Context for Long-Duration Engineering Tasks

FACT STATEMENT

Zhipu released GLM-5.2 in June 2026, open-sourcing weights under the MIT license and providing a 1M context version.

What happened

GLM-5.2 no longer treats long context merely as a capacity parameter; instead, it trains and evaluates around hour-level coding, research, and system optimization, further strengthening Zhipu's open agent engineering approach.

Technical significance

The model introduces IndexShare to reduce sparse attention index computation and improves MTP speculative decoding; the official release also discloses long-duration RL, anti-cheating, and ultra-long context inference service design.

Industry impact

Open models are beginning to publicly share details on long-duration training, evaluation anti-cheating, and inference services; competition is shifting from accepting more tokens to maintaining quality and cost control in complex trajectories.

Decision value

Long context is suitable for large codebases and multi-document tasks, but retrieval, compression, and phased acceptance should be used to control invalid input and tail costs.

What to watch

Observe quality degradation under 1M context, KV cache cost, long-task failure rate, community reproduction, and production throughput on different hardware.

DECISION BRIEF

Turn the evidence into a decision.

See how AIGC.NEWS separates verified change, judgment, and the next signal to watch.